Efficiency
and effectiveness are two of the most important properties of Inductive
Logic Programming (ILP) systems. This thesis presents a new efficient
and effective formalism for ILP, called Pattern Based Inductive Logic
Programming (PB-ILP). This formalism is based on the concepts of
instances and patterns. While an instance represents a specific case of
the concept to be learned, a pattern represents the structural
information of a set of instances with the same properties.
In
this formalism, the instances and patterns are first obtained, and then
a theory about the target concept is constructed from the obtained
patterns. While inducing rules from the obtained patterns, we employ a
new search algorithm, where new search heuristics are used, which
require only negative coverage information. This search algorithm
significantly reduces the total coverage test cost and provides new
opportunities for learning different theories when multiple solutions
exist.
The PB-ILP formalism has been implemented in a new ILP
system—the Pattern-based Induction Engine (PIE). This system is shown
to be more efficient and effective than the state of the art ILP
system, ALEPH, over a significant set of benchmarks.
Semantic Web Language Layering
with Ontologies, Rules, and Meta-Modeling
Jos de Bruijn
University of Innsbruck
Austria
The
recent advent of the Semantic Web has increased the interest in the use
of formal Knowledge Representation (KR) languages in order to allow
automated processing of, and reasoning with, information on the Web. A
prominent KR paradigm on the Web is the family of Description Logics
(DLs), which are subsets of classical First-Order Logic (FOL); the W3C
Web Ontology Language OWL, and especially its sub-language OWL DL, is
strongly related to DL. Another KR paradigm receiving widespread
attention in this context is that of rules, in the form of Logic
Programming (LP); it is used in, for example, Semantic Web policies and
Semantic Web services. Meta-modeling is a feature many deem useful, or
even necessary, for the Semantic Web, and is present in the Semantic
Web languages RDFS and OWL Full. F-Logic is a formalism that allows
meta-modeling, in the spirit of RDFS, but is in many respects more in
line with standard KR and database languages.
It is the goal of
this thesis to combine DL ontologies, LP rules, and meta-modeling
capabilities in a single unifying language framework for the Semantic
Web.
We propose WSML as such a unifying framework. The WSML-DL
sub-language corresponds to the DL SHIQ, and the WSML-Rule sub-language
corresponds to LP with negation, extended with F-Logic-based
meta-modeling. Interoperation between these sub-languages can be
achieved through a common subset (WSML-Core) or a common superset
(WSML-Full). In the technical design of the subset and the superset we
face two major challenges: the combination of DL-style and
F-Logic-style ontology modeling, and the interoperation between
ontologies based on classical FOL (e.g., DL) and rules based on
nonmonotonic LP.
- We address the combination of DL-style and
F-Logic-style ontology modeling by defining a straightforward
translation from FOL to F-Logic, and show that this translation
preserves validity for the class of cardinal formulas. We exhibit a
novel class of cardinal formulas that includes the DL SHIQ. We use this
result to show that the translation from SHIQ to F-Logic preserves
entailment.
- We address the interoperation between FOL and LP by
analyzing a number of representational issues that occur when combining
FOL theories with logic programs, and present a novel approach to
combining FOL and LP using first-order autoepistemic logic, an
expressive nonmonotonic logic, as a unifying formalism that features a
tight integration between the rules and the ontology. We show how
different embeddings of logic programs lead to difference semantics for
the combination.
Using these technical results, we specify a
semantic framework for WSML, define WSML-Core and WSML-Full, and
discuss language layering in WSML.
To address the relationship
with the basic Semantic Web language RDF(S) we define embeddings of the
language in F-Logic, and demonstrating how standard KR techniques can
be used for reasoning with and extending RDFS. Finally, we consider
intensional OWL, a variant of OWL Full more suitable for rule-based
processing and extension.